Pooled Standard Deviation Calculator — combine two or more groups

The pooled standard deviation calculator above takes the sample size and spread of two, three or four groups and merges them into a single estimate of the common population standard deviation. It weights each group by its degrees of freedom rather than by its raw sample size, returns the pooled variance and the pooled standard deviation, reports the total degrees of freedom that a downstream test would use, and flags how far apart the group variances actually are. That last number matters more than most textbooks admit, because the whole procedure rests on an assumption that the groups share one underlying spread.

Arb Digital publishes this page because pooled standard deviation is the quiet input behind a long list of familiar results. It sits inside the independent-samples t-test, inside Cohen's d, inside the residual term of a one-way ANOVA, and inside the margin of error of any confidence interval built around a difference between two means. People quote those outputs constantly without ever seeing the pooling step that produced them. Exposing the arithmetic makes it obvious why two groups of unequal size do not contribute equally, and why an outlying group with a huge variance can quietly wreck an estimate that looks perfectly respectable.

What This Pooled Standard Deviation Calculator Does

It computes a weighted average of the group variances, using each group's degrees of freedom as the weight, then takes the square root. That is the only correct way to combine spreads. Averaging the standard deviations directly is wrong, and averaging the variances with equal weights is wrong whenever the groups differ in size. The tool also derives the standard error of the difference between two means, which is the quantity a t-test divides by, and it exposes the ratio of the largest group variance to the smallest as a diagnostic.

This is a narrower job than our standard deviation calculator, which takes a list of raw values and describes one sample. Use that page when you have data; use this one when you already have summary statistics for several groups and need a single combined spread. It is also distinct from the standard error calculator, which converts one sample's SD into the standard error of its own mean rather than combining several samples. If you want the underlying squared quantity for a single dataset, the variance calculator handles that directly.

How to Use It

1. Say what the second column holds. Standard deviation is the default. Switch to variance if your output table reports s squared, because entering an SD where a variance is expected will inflate the pooled figure by roughly the square of the true value.
2. Enter each group's sample size. Use the number of observations, not the degrees of freedom. The calculator subtracts one per group itself.
3. Enter each group's spread. Use the sample standard deviation with the n minus 1 denominator, which is what every statistics package reports by default.
4. Set unused groups to a sample size of 0. Two groups is the common case; three and four are there for ANOVA-style designs.
5. Read the variance ratio before the pooled SD. If the groups plainly do not share a spread, the headline number is an average of things that should not have been averaged.

The Formula and How It's Calculated

For k groups with sizes n₁ through n_k and sample variances s₁² through s_k², the pooled variance is the sum of (n_i − 1)s_i² divided by the sum of (n_i − 1). The pooled standard deviation is the square root of that. The denominator, the total sample size minus the number of groups, is the degrees of freedom carried into whatever test comes next.

Take the default values. Group one has 12 observations with an SD of 4.5, so its variance is 20.25 and its weight is 11. Group two has 15 observations with an SD of 5.2, so its variance is 27.04 and its weight is 14. The weighted sum is 11 × 20.25 + 14 × 27.04 = 222.75 + 378.56 = 601.31. Divide by 11 + 14 = 25 and the pooled variance is 24.0524. The square root gives a pooled standard deviation of 4.9043 on 25 degrees of freedom. Notice that the answer sits closer to 5.2 than to 4.5, because the larger group carries more weight — the plain average of the two SDs, 4.85, is not the same number and is not the right one.

The standard error of the difference between the two group means follows as s_p × √(1/n₁ + 1/n₂), which here is 4.9043 × √(1/12 + 1/15) = 4.9043 × 0.3873 = 1.8994. The NIST/SEMATECH e-Handbook section on comparing two process means sets out the identical formula and shows where it enters the test statistic.

The Equal-Variance Assumption, and What It Costs

Pooling is only defensible if the groups are samples from populations with the same standard deviation. That is a genuine assumption about the world, not a formality. When it holds, pooling is the better estimator: you are using every observation to estimate one quantity, so the estimate is more stable and the test has more degrees of freedom and more power. When it fails, pooling produces a number that describes none of the groups and a standard error that is systematically wrong.

The direction of the error depends on which group is larger. If the bigger sample happens to be the one with the smaller variance, the pooled SD is dragged down, the standard error is understated, and the test becomes too willing to declare a difference. If the bigger sample has the larger variance, the opposite happens and the test becomes too conservative. This is why unequal variances combined with unequal sample sizes is the genuinely dangerous case, while unequal variances with equal sample sizes is comparatively forgiving.

The variance ratio in the results grid is a rough screen for this. A ratio near 1 is comfortable. A ratio around 2 or 3 is usually tolerable if the group sizes are similar. Beyond about 4, with unbalanced groups, the pooled result should not be trusted, and the Welch approach — which never pools and adjusts the degrees of freedom instead — is the standard alternative. Formal tests for equal variance exist, including Levene's test and Bartlett's test, but they are themselves sensitive to non-normality, so the ratio and a look at the data usually tell you more.

Why It Is Not the Standard Deviation of the Combined Data

This is the single most common misreading of the output. The pooled standard deviation measures spread within groups. It deliberately ignores any difference between the group means. If you stacked all the observations into one pile and computed an ordinary standard deviation, you would get a larger number whenever the group means differ, because that calculation also picks up the spread between groups.

An example makes the gap obvious. Suppose two groups each have 10 observations and an SD of 3, but one group's mean is 100 and the other's is 120. The pooled SD is exactly 3, because within each group the spread is 3. The SD of the combined 20 numbers is well over 10, because the two clusters sit 20 units apart. Both numbers are correct answers to different questions. Pooled SD answers "how noisy is a single observation once I account for which group it came from"; the combined SD answers "how variable is the whole collection". Reporting one when the reader expects the other is a real error, and it happens most often when someone tries to summarise several study arms in a single line.

Where the Pooled SD Turns Up Downstream

The independent-samples t-test divides the difference in means by exactly the standard error the grid reports, then compares the result against a t distribution with the pooled degrees of freedom. Our t-test calculator runs that comparison end to end. Cohen's d divides the same difference by the pooled SD instead of by the standard error, which is why the effect size calculator asks for group SDs and pools them internally — effect size is a difference expressed in standard deviations, and the pooled SD is the ruler.

One-way ANOVA generalises the idea to more than two groups. The mean square error in an ANOVA table is precisely the pooled variance from this page, and its degrees of freedom are the same total-minus-k figure. That is why an ANOVA on two groups gives an F statistic exactly equal to the square of the t statistic from a pooled t-test: they are the same calculation wearing different clothes. The degrees of freedom calculator covers how that count changes across designs, and a confidence interval calculator will use the same pooled standard error to build an interval around the difference.

Three or More Groups, and Groups of Size One

Nothing about the formula changes when you add groups; you simply add more terms to both sums. What does change is how much a single badly behaved group can distort the result, because the pooled estimate is now supposed to represent a common spread across more populations. Screening the variance ratio matters more with four groups than with two, since one group with a much wider spread will pull the pooled figure up while every other group's residuals are quietly overstated by it.

Groups with exactly one observation contribute nothing. Their weight, n minus 1, is zero, so they add zero to the numerator and zero to the denominator. The calculator handles this by ignoring them entirely, which is correct — a single observation carries no information about spread. Groups of size zero are ignored the same way. If every group has one or zero observations, there is nothing to pool and the tool reports no result rather than a misleading zero. Penn State's STAT 500 Applied Statistics notes work through the two-sample and multi-group cases with the same weighting scheme.

Common Mistakes to Avoid

• Averaging the standard deviations directly — spreads combine through their squares, so the plain average of two SDs is not the pooled SD unless the groups happen to be identical in size and spread.
• Weighting by n instead of n minus 1 — the weights are degrees of freedom. Using raw sample sizes gives a slightly biased estimate that gets worse as the groups get smaller.
• Feeding variances into the standard deviation field — a variance of 27 entered as an SD produces a pooled figure roughly five times too large. Use the selector at the top.
• Pooling wildly unequal variances anyway — when the ratio is large and the group sizes are unbalanced, the pooled standard error is wrong in a predictable direction and the resulting p-value is not trustworthy.
• Reporting the pooled SD as the spread of the combined dataset — it excludes the variation between group means, so it will always be the smaller of the two whenever those means differ.

Related Free Tools From Arb Digital

Describe a single dataset with the descriptive statistics calculator, work out how many observations you need before you collect them with the sample size calculator, convert an individual value into a standardised distance with the z-score calculator, or check whether an observed gap clears the significance bar with the p-value calculator. The full free online tools hub lists every statistics calculator we publish.

Frequently Asked Questions

This page explains a statistical calculation for educational purposes only. It is not medical, legal, or financial advice, and results should be interpreted alongside the assumptions of whichever test you are running.